This dataset contains a slice of 200 samples from the AG News dataset (test split). The picked 200 samples are potential misclassifications of the original test data.
Approach
- Fine-tune DistilBERT with 10k samples from the training data (out of 120k)
- Do a forward pass with the model, storing the loss
- Sort the samples based on the loss
This is a repository for demonstration purposes